Home › AWS › AWS data services › IoT SiteWise
AWS data services · Ingestion and CDC · about 12 min read
AWS IoT SiteWise and SiteWise Edge
Collects readings from plant machines through an on-site gateway, and optionally models and stores them as industrial assets in the cloud.
In plain words
What it is
A tyre or food plant is full of machines that speak industrial protocols, not web protocols. A curing press reports its temperature and pressure through OPC UA (the common standard for reading tags from machine controllers), an older mixer may only speak Modbus, and a plant historian keeps years of readings in its own format. None of these can post a JSON message to a cloud service on their own. Something on the plant network has to read the tags, stamp them with time and quality, and carry them out.
AWS IoT SiteWise is the AWS service built for that job, and it has two halves. SiteWise Edge is software that runs on a small server inside the plant, called a gateway. It reads OPC UA and Modbus tags, keeps a local buffer on disk when the internet link drops, and forwards the readings when the link is back. The cloud half of SiteWise lets you describe your plant as a tree of assets (plant, line, press) with properties (temperature, pressure, cycle count), stores the readings in its own hot, warm and cold tiers, and computes metrics such as hourly averages or overall equipment effectiveness.
Take one example. A plant has 40 curing presses, each with 25 tags read once a second, so about 1,000 readings a second. SiteWise Edge on one gateway reads all of them through OPC UA. On this site's reference platform, the gateway sends those readings on to Amazon Kinesis Data Streams, where Databricks lands them in a bronze table every 15 minutes and a live lane watches them for alerts. The cloud asset model is optional and, on this platform, skipped, because Unity Catalog and the silver and gold tables already describe presses and lines.
A translator stationed on the factory floor who understands every machine's dialect, writes down each reading in one common language, and keeps a notebook of everything said while the phone line to head office is down.
Big picture
Where it fits
The lit lane is where IoT SiteWise works. Boxes on the left feed it; boxes on the right use what it produces.
SiteWise sits at the very first hop of the streaming lane: between the machines and the stream belt. On the reference platform the plant gateway runs SiteWise Edge on AWS IoT Greengrass (the AWS runtime for software on edge devices). The collectors read the presses and mixers, and the readings leave the plant for Kinesis Data Streams in the home Region. From there, two readers use the same events: Managed Service for Apache Flink for seconds-level alerts, and a Databricks streaming table that writes bronze and then silver and gold.
There is one wrinkle to know before you build. The older gateway type, called Classic streams or V2, moves data through Greengrass stream manager, which can export straight to Kinesis Data Streams or S3. The newer MQTT-enabled V3 gateway that AWS now recommends for self-hosted installs uses destinations, and its built-in destinations are SiteWise real time and SiteWise buffered to S3 as Parquet files. To reach Kinesis from a V3 gateway you add a small Greengrass component that subscribes to the local MQTT topics and writes to Kinesis, or you land Parquet in S3 and let Auto Loader pick it up. Both keep the site's rule: data leaves the plant once, into storage we own.
What SiteWise is not, on this platform, is the safety system. Machine alarms and interlocks stay in the machines' own controls and in SiteWise Edge at the plant. The platform is for analysis and decisions, not for stopping a press. CloudWatch watches the gateway and the Kinesis delay, and the mover health table in the ops catalog shows both next to the Databricks runs.
How it works
The ideas you need
| Idea | What it means |
|---|---|
| Gateway | The server at the plant that runs SiteWise Edge. It reads machines, buffers to disk and forwards data. It runs on AWS IoT Greengrass V2, or on Siemens Industrial Edge. |
| Data stream | One time series in SiteWise, usually one machine tag such as /plant1/press07/temperature. Each reading is a timestamp, a quality flag and a value, known as a TQV. |
| Asset model and asset | A template (Curing press, with properties temperature and pressure) and its instances (Press 07). This is the cloud modelling part the reference platform skips. |
| Measurement, transform and metric | Raw readings are measurements. A transform changes one value, such as Fahrenheit to Celsius. A metric aggregates over time, such as a five-minute average. SiteWise bills each computation. |
| Destination and path filter | On V3 gateways, a destination says where data goes (SiteWise real time or S3 buffered). A path filter is an MQTT topic pattern that picks which streams go there. |
| Classic streams, V2 gateway | The older gateway type that moves data through Greengrass stream manager. Its publisher can export to Kinesis or S3, which is the route the site's design names. |
| Storage tiers | Hot storage for recent data, warm storage for older queryable data, and a cold tier written as files to your own S3 bucket. |
| Data quality flag | GOOD, BAD or UNCERTAIN on each reading, passed from OPC UA. Keep it in bronze; silver rules decide what to do with BAD readings. |
Choosing
When to use it, and when not
Use it when
- Your plant machines speak OPC UA or Modbus and you need a supported collector rather than a home-built Python service.
- The plant link is unreliable and you need the gateway to buffer readings on local disk and resend them in order when the link returns.
- You run many plants and want AWS to deploy and update the collector software centrally through Greengrass.
- Operators need a simple asset-based view of live machine values and you accept paying per data point for SiteWise cloud storage on a small, chosen set of tags.
Avoid it when
- The plant already runs Kepware, Ignition or another industrial platform that can publish MQTT or write to Kinesis; adding a second collector adds a runtime without adding value.
- You plan to use SiteWise cloud storage as the analytics store for every tag; per data point pricing grows with volume, and the lakehouse already keeps history more cheaply.
- You need a safety interlock or a sub-second machine stop; that belongs in the machine controller, not in a cloud service.
- The devices are connected products in the field, such as sensors on vehicles; they need per-device identity through IoT Core, not a plant gateway.
Production grade
What goes wrong in production
| Problem | What breaks | What to do |
|---|---|---|
| Gateway type confusion | Teams follow an older guide written for V2 Classic streams with a Kinesis export, then build a V3 gateway whose built-in destinations are only SiteWise and S3. | Decide the gateway type first. With V3, either land Parquet in S3 for Auto Loader or add a small Greengrass component that writes to Kinesis. Write the choice into the binding file. |
| Paying twice for history | Readings go to SiteWise hot storage and also to Kinesis and the lake. The SiteWise bill grows per data point with no extra use. | Turn off the SiteWise real-time destination with a path filter, or point it at a short list of tags operators actually view. |
| Clock and quality drift | Plant controllers have wrong clocks or send BAD quality readings, so silver joins and windows go wrong. | Keep both device time and gateway time in the envelope, keep the quality flag, sync plant clocks with NTP, and flag large gaps in a silver check. |
| Buffer fills during a long outage | A multi-day link outage fills the gateway disk; old data is dropped by the configured policy. | Size the gateway disk for the longest outage you accept, set the delete-when-full policy deliberately, and alarm on disk use. |
| Tag name chaos | Each plant names tags differently, so one press metric has five names across sites. | Agree a topic hierarchy such as plant/line/machine/tag before the first plant, and map tags to it in the collector, not in Spark. |
| Monitor portals on a dead end | Teams build operator dashboards in SiteWise Monitor, which is now in maintenance mode and closed to new customers. | Build new dashboards in Amazon Managed Grafana with the SiteWise plugin, or in Databricks AI/BI on gold tables. |
| Alarms tied to IoT Events | SiteWise alarms were built on AWS IoT Events, which reached end of support on 20 May 2026. | Follow the AWS migration guide to CloudWatch alarms for SiteWise properties, or move alerting to the Flink live lane. |
Alternatives
What to compare it with
| Alternative | Pick the alternative when | Pick IoT SiteWise when |
|---|---|---|
| Kepware, Ignition or other industrial software | The plant already licenses one and its team knows it; it can publish MQTT or write to Kinesis. | You have nothing at the plant yet and want AWS to manage the collector and its updates. |
| Our own Python collector on a small box | Never, for production; the site drops this option. | Always: you would own buffering, retries, updates and security patches. |
| AWS IoT Core directly from machines | Each device needs its own identity, commands back, or shadows, as with connected tyres in the field. | Plant machines can share one gateway; per-message pricing would add about a thousand dollars a month at the site's scale. |
| Azure IoT Operations | The platform runs on Azure; it is the Microsoft route for OPC UA at the edge into Event Hubs or Fabric. | The platform is on AWS; keep one cloud's edge runtime. |
| Databricks Zerobus Ingest | A single app writes events straight into one Delta table and nothing else needs them. | Two readers need the same plant readings, a live lane and the lake, with replay from the log. |
Cost
How you pay
SiteWise bills several meters separately. Ingestion into SiteWise is charged per message, metered in small blocks of bytes or data points per stream for real-time ingestion and in larger blocks for buffered ingestion. Retrieval is charged when you query or publish values. Transforms, metrics and computed aggregates are charged per computation, and queries by data scanned. Hot and warm storage are charged by volume over time, and the cold tier is charged per S3 object written plus your normal S3 costs. At the edge, the data collection pack is free and the data processing pack is charged per active gateway per month. Greengrass itself is billed separately per active core device. SiteWise Monitor, for existing users, is billed per active user. Check the current pricing page for your Region before estimating.
What drives the bill
- The number of data points sent to SiteWise cloud ingestion, which grows with tags multiplied by reading frequency.
- How many transforms and metrics you define, because each computed value is billed.
- How long data stays in hot and warm storage before it moves to cold storage or is deleted.
- The number of active gateways, Greengrass core devices and, if used, the data processing pack.
Cost traps
- Leaving the default V3 real-time destination path filter set to everything, which sends every tag to SiteWise hot storage as well as to the lake.
- Defining many fine-grained metrics on every asset, which multiplies computations every minute whether anyone looks or not.
- Querying SiteWise history from dashboards with interpolation, which meters every stored value read.
- Counting only SiteWise and forgetting Greengrass, Kinesis, S3 requests and the gateway hardware in the total.
Prices change; check the AWS pricing page for your region before deciding.
Security and governance
Keeping it safe
- The gateway gets AWS credentials through the Greengrass token exchange service, which swaps the device certificate for a short-lived IAM role. Scope that role to the one Kinesis stream or S3 prefix it writes to.
- OPC UA connections should use Sign and Encrypt security mode with trusted certificates; store any OPC UA user names and passwords in AWS Secrets Manager, which the collector can read.
- Put the gateway in a plant network zone that can reach the machines and only outbound HTTPS and MQTT to AWS; no inbound ports from the internet. Over Direct Connect or VPN, use VPC endpoints for Kinesis, S3 and STS.
- SiteWise encrypts data at rest by default and can use a customer managed AWS KMS key; S3 buffered files should land in the landing bucket encrypted with the catalog's KMS key.
- CloudTrail records SiteWise and Greengrass control-plane calls such as gateway configuration changes and deployments; keep them in the audit trail.
- Plant readings rarely hold personal data, but operator IDs or badge numbers sometimes appear in tags; tokenise them in the lander like any other personal field.
Operating it
What to watch, and the limits
| Watch | Why |
|---|---|
| Gateway heartbeat and connectivity | SiteWise reports the gateway's last contact; alarm if a plant gateway has not checked in for a few minutes. |
| Greengrass component health | Each collector and publisher component reports a state such as RUNNING or BROKEN in Greengrass; alarm on anything not running. |
| Local buffer and disk use | How full the gateway's buffer disk is; a rising value means data is not leaving the plant. |
| Kinesis IncomingRecords per plant | Readings arriving on the stream; a drop to zero for one plant points at that gateway. |
| SiteWise ingestion throttling | Throttled BatchPutAssetPropertyValue calls in CloudWatch mean the account rate quota is too low. |
| Reading lag in bronze | Bronze ingest time minus device time per plant; it shows clock drift and buffering delays in one number. |
| Limit | Value and notes |
|---|---|
| Gateways per Region per account | 100 by default, adjustable |
| Data points ingested per second per account | 5,000 by default, adjustable; this applies to SiteWise cloud ingestion, not to Kinesis |
| Entries per second per asset property | 10, not adjustable |
| OPC UA sources per gateway | 100, not adjustable |
| Path filters per gateway | 20,000 across all destinations; one real-time destination per gateway |
| Properties per asset model | 5,000, adjustable |
Limits and quotas checked 2026-11-11; many can be raised through Service Quotas.
Decisions
Situations and verdicts
| Situation | Choice | Verdict |
|---|---|---|
| Plant with 40 presses on OPC UA, 1,000 readings a second | SiteWise Edge on Greengrass, readings to Kinesis | Pursue: supported collectors, offline buffering, central updates; data leaves once into storage we own. |
| Store all plant history in SiteWise warm storage | SiteWise cloud storage as the analytics store | Drop: per data point pricing duplicates the lakehouse, which keeps history in Delta at S3 prices. |
| New operator dashboards on live machine values | SiteWise Monitor portals | Drop: Monitor is closed to new customers; use Managed Grafana with the SiteWise plugin or Databricks dashboards. |
| Existing SiteWise alarms on press temperature | Keep SiteWise asset models for those few tags, move alarms to CloudWatch | Keep if you have it: the alarms must move off IoT Events, but the asset model can stay for the tags that need it. |
FAQ
Questions people ask
What is the difference between AWS IoT SiteWise and SiteWise Edge?
SiteWise Edge is the software that runs on a gateway inside the plant to read machines and buffer data. SiteWise in the cloud models assets, stores time series and computes metrics. You can use the edge part and send data elsewhere without using the cloud storage.
Can SiteWise Edge send data to Kinesis Data Streams?
Classic streams V2 gateways can export through Greengrass stream manager to Kinesis or S3. The newer MQTT-enabled V3 gateways have built-in destinations for SiteWise real time and for S3 as Parquet; to reach Kinesis you add a Greengrass component that reads the local MQTT topics and writes to Kinesis.
Is AWS IoT SiteWise being retired?
No. SiteWise is generally available. Only the SiteWise Monitor portal feature is in maintenance mode and closed to new customers, and SiteWise alarms built on IoT Events must move because IoT Events reached end of support on 20 May 2026.
Does SiteWise replace a plant historian?
It can collect from one and keep history in the cloud, but on a lakehouse platform the history lives in Delta tables. Keep the plant historian for local operations and treat SiteWise Edge as the collector.
How do I read SiteWise data in Databricks?
The cheapest path is to not read it from SiteWise at all: send readings to Kinesis or S3 from the gateway and read them with a streaming table or Auto Loader. If data already sits in SiteWise, its cold tier writes files to your S3 bucket, which Databricks can read.
Sources
Checked against
- AWS docs: SiteWise Edge destinations and path filters
- AWS docs: MQTT-enabled V3 gateways
- AWS docs: SiteWise Monitor availability change
- AWS General Reference: IoT SiteWise endpoints and quotas
- AWS IoT SiteWise pricing
Checked 2026-11-11.
Related
Where it connects
Streaming on AWS and Databricks · Data platform on AWS + Databricks · AWS IoT Greengrass · Amazon Kinesis Data Streams · AWS IoT Events and IoT Analytics · Amazon Managed Grafana · All AWS data services